Pith. sign in

REVIEW 2 major objections 4 minor 35 references

Optimal-transport flows make Bayesian optimization of wind-farm layouts both better and twice as fast by removing artificial permutation copies.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 17:27 UTC pith:I66FXBKO

load-bearing objection Clean OT-flow construction for continuous layout BO that beats vanilla BO on a five-turbine wind task and halves wall-clock; the ranking is measured only on a sorted XGBoost ensemble, so transfer to real FLORIS is still open. the 2 major comments →

arxiv 2606.00009 v1 pith:I66FXBKO submitted 2026-03-27 cs.AI

Optimal Transport-based Permutation-Invariant Bayesian Optimization of Offshore Wind Farm Layouts

Antonio Candelieri , Laurens Bliek This is my paper
classification cs.AI
keywords Bayesian optimizationpermutation invarianceoptimal transportoffshore wind farm layoutlayouts versus point-cloudsWasserstein distanceGaussian process
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Many placement problems treat identical objects as ordered lists even though swapping any pair leaves the objective unchanged. That artificial ordering multiplies every true layout by m! replicas, turning a smooth landscape into a forest of needles that standard Bayesian optimization must rediscover. The paper shows that the remedy is to fix a reference cloud outside the domain and to work exclusively in the space of optimal-transport flows that map the reference onto candidate layouts. Because those flows are unique almost surely, the Gaussian process never sees the replicas. On a five-turbine offshore wind-farm task the resulting algorithm, PIBO, consistently finds higher annual-energy-production layouts than vanilla BO in either the physical plane or the unrestricted flow space, while the total wall-clock time drops by roughly half. The same device therefore both improves solution quality and simplifies the surrogate model that drives the search.

Core claim

Performing Bayesian optimization over optimal-transport flows from a fixed exterior reference cloud removes the combinatorial replication that otherwise multiplies every layout by m!, yielding higher annual-energy-production placements than vanilla BO (in either physical coordinates or unrestricted flows) and cutting total computation time roughly in half on the five-turbine offshore wind-farm problem.

What carries the argument

The optimal-transport flow X* = (T*^T - I)P-bar that uniquely (almost surely) maps a fixed exterior reference cloud onto any candidate layout; Bayesian optimization is performed entirely in the space of these flows, so the Gaussian process never sees permutation replicas.

Load-bearing premise

All performance claims rest on a median-of-five XGBoost ensemble that was trained on sixty thousand FLORIS evaluations and then artificially noised; superiority is measured only against this surrogate, not against the original expensive wake simulator.

What would settle it

Re-run the identical experimental protocol (same seeds, same budget, same acquisition settings) on the true FLORIS simulator instead of the XGBoost ensemble and check whether PIBO still produces statistically higher AEP and still finishes in roughly half the wall-clock time of the two vanilla BO baselines.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes PIBO, a Bayesian optimization method for continuous layout problems that are invariant to permutations of the points. It maps candidate layouts to optimal-transport flows from a fixed reference point-cloud (via the Kantorovich formulation of the 2-Wasserstein problem), fits a GP and optimizes GP-UCB in that flow space, then projects each new query back onto an optimal flow before evaluation. The method is evaluated on a five-turbine offshore wind-farm layout task whose objective is annual energy production subject to a minimum-separation constraint. Against LHS, URS, TPE and two vanilla-BO baselines (physical space and unrestricted flow space), PIBO with an Exponential kernel and ξ=6 yields the highest average AEP while roughly halving wall-clock time relative to the other BO variants. Code is released.

Significance. Permutation-invariant continuous layout optimization is a recurring industrial setting (sensor placement, well placement, wind farms) for which standard BO is known to suffer from an m!-fold replication of every local and global optimum. Framing the search space as OT-optimal flows is a clean, theoretically motivated remedy that simultaneously removes the replication and simplifies the GP fit. The empirical ablation (physical space → unrestricted flows → optimal flows) isolates the contribution of the OT projection, the shared-initialization design with 30 seeds is careful, and the public code release supports reproducibility. If the superiority transfers beyond the surrogate used here, the technique would be immediately useful for a class of expensive black-box layout problems.

major comments (2)
  1. All superiority claims (Tables 1–2, Figures 2–4, wall-clock comparison) are measured exclusively against the median-of-five XGBoost ensemble of Appendix A, which was itself trained after sorting layouts by horizontal (then vertical) coordinate and then artificially noised. Because that preprocessing already injects a crude form of permutation invariance, the ranking advantage of OT-optimal flows over unrestricted flows or physical-space vectors could be an artifact of the surrogate rather than a property that would hold for the true FLORIS wake simulator. No final layout is re-evaluated with FLORIS; without at least a modest set of such re-evaluations the central claim remains conditional on surrogate fidelity.
  2. Acquisition-function optimization is performed by a fixed 10 000-point LHS (Section 4.1). For a 10-dimensional continuous space this is a very coarse maximizer; the reported performance gaps could therefore be partly attributable to how well each search space is explored by the same crude optimizer rather than to the intrinsic difficulty of the landscapes. A short sensitivity study (or a more accurate AF optimizer) would strengthen the ablation.
minor comments (4)
  1. Proposition 1 (Section 3) rewrites the separation constraint under the assumption that the reference points already violate it; the proof sketch is terse and the intermediate inequality (15)–(17) would benefit from a short geometric interpretation.
  2. Algorithm 1 output line reports y+ = min{…} while the problem is a maximization; the typo is harmless but should be corrected.
  3. Figure 1 caption and surrounding text speak of “four global optima” for the mirrored Bird function; a brief remark that two of them are exact replicas would avoid any impression that the landscape has been fundamentally altered.
  4. The reference-cloud sampling distribution (mean and covariance) is left unspecified beyond “outside the domain”; a concrete recipe would aid exact reproduction.

Circularity Check

0 steps flagged

No circularity: algorithmic method (OT flows for PI-BO) is independently defined and empirically ablated against baselines on a fixed surrogate; no fitted quantities re-presented as predictions.

full rationale

The paper's central claim is an algorithmic construction (PIBO: GP-UCB over optimal-transport flows from a fixed external reference cloud, Algorithm 1) plus an empirical head-to-head on a five-turbine layout task. OT theory (Monge/Kantorovich, Brenier remarks) is standard and applied without self-referential redefinition; Proposition 1 is a short algebraic rewrite of the distance constraint under the reference-cloud assumption and does not feed back into the objective. Performance numbers (Tables 1-2, Figs. 2-4) are obtained by running the same budget of 500 evaluations from shared LHS seeds against external baselines (LHS/URS/TPE/vanilla BO in physical and unrestricted-flow spaces). The only self-citation is to the EXPObench library and its FLORIS-derived data set (used solely as the source of the XGBoost ensemble surrogate in Appendix A); that citation supplies a fixed black-box evaluator and does not underwrite uniqueness, force the ranking, or close any derivation loop. No parameter is fitted to a subset of the reported results and then called a prediction; the surrogate itself is held fixed across all methods. Consequently the derivation chain contains no self-definitional, fitted-as-prediction, or load-bearing self-citation steps.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 1 invented entities

The central empirical claim rests on standard BO and OT machinery plus a handful of modeling choices (reference distribution, surrogate ensemble, inexact acquisition) that are not derived from first principles. No new physical entities are postulated.

free parameters (5)
  • GP-UCB exploration bonus ξ = 6 (selected)
    Hand-chosen discrete set {0,1,3,6}; best reported result uses ξ=6 with exponential kernel.
  • minimum turbine separation ρ = 0.1512
    Fixed at twice rotor diameter after rescaling; value 0.1512 in unit square.
  • reference cloud distribution parameters
    Mean and covariance of the multivariate normal used to sample P-bar outside the domain; chosen so that cost matrices are in general position.
  • ensemble noise scale σ = 1
    Uniform noise multiplier set to 1 in Appendix A to mimic simulator stochasticity.
  • number of LHS points for acquisition optimization = 10000
    Fixed at 10 000; inexact optimizer hyper-parameter.
axioms (3)
  • domain assumption When transport cost is convex and points are not on a fixed grid, the cost matrix is almost surely in general position, guaranteeing uniqueness of the optimal plan.
    Invoked in Section 2.2 to justify that the OT map is unique for almost all layouts; used to claim that the flow space is free of replications.
  • ad hoc to paper The median-of-five XGBoost ensemble plus one-sided noise is an adequate ranking surrogate for the true FLORIS wake simulator.
    All numerical claims rest on this surrogate (Appendix A); no direct validation against held-out FLORIS runs is reported for the final layouts.
  • domain assumption Gaussian-process UCB with exponential or squared-exponential kernel is a suitable acquisition strategy for the flow space.
    Standard BO assumption; no theoretical regret guarantee is claimed for the OT-transformed space.
invented entities (1)
  • optimal-flow search space for layouts independent evidence
    purpose: Represents each layout uniquely as the OT displacement field from a fixed external reference cloud, thereby removing permutation replicas from the GP domain.
    Core modeling device of PIBO; independent evidence is the empirical ablation showing improved sample efficiency and GP-fitting speed.

pith-pipeline@v1.1.0-grok45 · 18884 in / 2681 out tokens · 31140 ms · 2026-07-13T17:27:37.661050+00:00 · methodology

0 comments
read the original abstract

Bayesian Optimization (BO) is widely and successfully adopted for solving optimization problems having an expensive-to-evaluate, black-box, and non-convex objective function. However, the vanilla BO algorithm is not able to exploit possible symmetries characterizing the target problem. An intuitive case is given by optimal location problems, whose decision variables refer to a finite set of points within a continuous space, with the order of points not affecting the value of the objective function. We refer to this setting as optimization over layouts to distinguish from optimization over point-clouds where, instead, the order of points counts. As an instance of optimization over layouts we consider a real-life industrial-relevant application, that is the optimization of the layout of an offshore wind farm: given identical wind turbines, switching any pair of them has not any effect on the annual energy production. Based on Optimal Transport theory, we propose a Permutation-Invariant BO approach, namely PIBO, proved to provide better wind farm layouts when compared to the vanilla BO approach while cutting computation time roughly in half.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 2 canonical work pages

  1. [1]

    In: Zhang, Y., Hladik, M., Moosaei, H

    Hellan, S.P., Lucas, C.G., Goddard, N.H.: Bayesian optimisation against climate change: Applications and benchmarks. In: Zhang, Y., Hladik, M., Moosaei, H. (eds.) Learning and Intelligent Optimization, pp. 107–118. Springer, Cham (2026)

    2026

  2. [2]

    Applied Sciences (2024)

    Fotias, S.P., Ismail, I., Gaganis, V.I.: Optimization of well placement in carbon capture and storage (ccs): Bayesian optimization framework under permutation invariance. Applied Sciences (2024)

    2024

  3. [3]

    In: AAAI Conference on Artificial Intelligence (2021)

    Deshwal, A., Belakaria, S., Doppa, J.R., Kim, D.H.: Bayesian optimization over permutation spaces. In: AAAI Conference on Artificial Intelligence (2021)

    2021

  4. [4]

    Applied Soft Computing147, 110744 (2023) https://doi.org/10.1016/j.asoc.2023

    Bliek, L., Guijt, A., Karlsson, R., Verwer, S., de Weerdt, M.: Benchmark- ing surrogate-based optimisation algorithms on expensive black-box functions. Applied Soft Computing147, 110744 (2023) https://doi.org/10.1016/j.asoc.2023. 110744

    work page doi:10.1016/j.asoc.2023 2023

  5. [5]

    Archetti, F., Candelieri, A.: Bayesian Optimization and Data Science vol. 849. Springer, Cham (2019) 17

    2019

  6. [6]

    Cambridge University Press, Cambridge (2023)

    Garnett, R.: Bayesian Optimization. Cambridge University Press, Cambridge (2023)

    2023

  7. [7]

    Journal of Machine Learning Research12(10) (2011)

    Bull, A.D.: Convergence rates of efficient global optimization algorithms. Journal of Machine Learning Research12(10) (2011)

    2011

  8. [8]

    arXiv preprint arXiv:1406.7758 (2014)

    Wang, Z., Freitas, N.: Theoretical analysis of bayesian optimisation with unknown gaussian process hyper-parameters. arXiv preprint arXiv:1406.7758 (2014)

    Pith/arXiv arXiv 2014

  9. [9]

    arXiv preprint arXiv:1612.03117 (2016)

    Wabersich, K.P., Toussaint, M.: Advancing bayesian optimization: The mixed-global-local (mgl) kernel and length-scale cool down. arXiv preprint arXiv:1612.03117 (2016)

    Pith/arXiv arXiv 2016

  10. [10]

    Journal of Machine Learning Research20(50), 1–24 (2019)

    Berkenkamp, F., Schoellig, A.P., Krause, A.: No-regret bayesian optimization with unknown hyperparameters. Journal of Machine Learning Research20(50), 1–24 (2019)

    2019

  11. [11]

    In: International Conference on Learning and Intelligent Optimization, pp

    Candelieri, A., Signori, E.: Mle-free gaussian process based bayesian optimization. In: International Conference on Learning and Intelligent Optimization, pp. 81–95 (2024). Springer

    2024

  12. [12]

    Proceedings of the Genetic and Evolutionary Computation Conference (2021)

    Irurozki, E., L´ opez-Ib´ a˜ nez, M.: Unbalanced mallows models for optimizing expensive black-box permutation problems. Proceedings of the Genetic and Evolutionary Computation Conference (2021)

    2021

  13. [13]

    In: EvoAppli- cations (2022)

    L’opez-Ib’anez, M., Chicano, F., Gil-Merino, R.: The asteroid routing problem: A benchmark for expensive black-box permutation optimization. In: EvoAppli- cations (2022)

    2022

  14. [14]

    Xie, Z., Chen, L.: From sorting algorithms to scalable kernels: Bayesian optimiza- tion in high-dimensional permutation spaces. (2025)

    2025

  15. [15]

    In: Neural Information Processing Systems (2018)

    Kandasamy, K., Neiswanger, W., Schneider, J.G., P´ oczos, B., Xing, E.P.: Neural architecture search with bayesian optimisation and optimal transport. In: Neural Information Processing Systems (2018)

    2018

  16. [16]

    In: International Symposium on Information Processing in Sensor Networks (2010)

    Garnett, R., Osborne, M.A., Roberts, S.J.: Bayesian optimization for sensor set selection. In: International Symposium on Information Processing in Sensor Networks (2010)

    2010

  17. [17]

    Machine Learning114(1), 22 (2025)

    Elaarabi, M., Borzacchiello, D., Bot, P.L., Guennec, Y.L., Comas-Cardona, S.: Adaptive parameters identification for nonlinear dynamics using deep permuta- tion invariant networks. Machine Learning114(1), 22 (2025)

    2025

  18. [18]

    Neural Computing and Applications37(5), 3505–3520 (2025) 18

    P´ erez-Mon, O., Moreo, A., Coz, J.J.d., Gonz´ alez, P.: Quantification using permutation-invariant networks based on histograms. Neural Computing and Applications37(5), 3505–3520 (2025) 18

    2025

  19. [19]

    arXiv preprint arXiv:2203.07546

    Balan, R., Haghani, N., Singh, M.: Permutation invariant representations with applications to graph deep learning (2022). arXiv preprint arXiv:2203.07546

    Pith/arXiv arXiv 2022

  20. [20]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    Liu, W., Zhu, Y., Zha, Y., Wu, Q., Jian, L., Liu, Z.: Rotation-and permutation- equivariant quantum graph neural network for 3d graph data. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    2025

  21. [21]

    Neurocomputing 625, 129577 (2025)

    Patil, S.S., Ramteke, M., Rathore, A.S.: Permutation invariant self-attention infused u-shaped transformer for medical image segmentation. Neurocomputing 625, 129577 (2025)

    2025

  22. [22]

    In: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp

    Selvaraj, E.K., Shah, A., Agrawal, S., Hedaoo, R., Golsefid, S.: Permutation- invariant transformers for attribute embeddings in information retrieval. In: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 3060–3064 (2025)

    2025

  23. [23]

    arXiv preprint arXiv:2601.19491 (2026)

    Chen, X., Zhao, S., Ma, F., Cheng, E., Burnett, I.S.: Permutation-invariant physics-informed neural network for region-to-region sound field reconstruction. arXiv preprint arXiv:2601.19491 (2026)

    arXiv 2026

  24. [24]

    Machine Learning110, 857–879 (2021)

    Kim, J., McCourt, M.J., You, T., Kim, S., Choi, S.: Bayesian optimization with approximate set kernels. Machine Learning110, 857–879 (2021)

    2021

  25. [25]

    Chapman and Hall/CRC, London (2020)

    Gramacy, R.B.: Surrogates: Gaussian Process Modeling, Design, and Optimiza- tion for the Applied Sciences. Chapman and Hall/CRC, London (2020)

    2020

  26. [26]

    Williams, C.K., Rasmussen, C.E.: Gaussian Processes for Machine Learning vol

  27. [27]

    MIT press, Cambridge, MA (2006)

    2006

  28. [28]

    Proceedings of the IEEE104(1), 148–175 (2015)

    Shahriari, B., Swersky, K., Wang, Z., Adams, R.P., De Freitas, N.: Taking the human out of the loop: A review of bayesian optimization. Proceedings of the IEEE104(1), 148–175 (2015)

    2015

  29. [29]

    Foundations and Trends®in Machine Learning11(5-6), 355–607 (2019)

    Peyr´ e, G., Cuturi, M.,et al.: Computational optimal transport: With applications to data science. Foundations and Trends®in Machine Learning11(5-6), 355–607 (2019)

    2019

  30. [30]

    Santambrogio, F.: Optimal transport for applied mathematicians (2015)

    2015

  31. [31]

    Kim, H., Liu, C., Chen, Y.: Bayesian optimization with inexact acquisition: is random grid search sufficient? arXiv preprint arXiv:2506.11831 (2025)

    Pith/arXiv arXiv 2025

  32. [32]

    In: International Conference on Machine Learning (2013)

    Bergstra, J., Yamins, D., Cox, D.D.: Making a science of model search: Hyper- parameter optimization in hundreds of dimensions for vision architectures. In: International Conference on Machine Learning (2013)

    2013

  33. [33]

    In: ACM SIGKDD, pp

    Chen, T., Guestrin, C.: XGBoost: A scalable tree boosting system. In: ACM SIGKDD, pp. 785–794 (2016) 19

    2016

  34. [34]

    Version 2

    Bliek, L., Guijt, A., Karlsson, R., Verwer, S., de Weerdt, M.: Raw data of the EXPensive Optimization benchmark library (EXPObench). Version 2. 4TU.ResearchData (2021). https://doi.org/10.4121/14247179.v2

    work page doi:10.4121/14247179.v2 2021

  35. [35]

    Version 2.1.1

    NREL: FLORIS. Version 2.1.1. GitHub (2020). https://github.com/NREL/floris 20 Appendix A Description of wind farm ensemble predictor The wind farm ensemble predictor consists of an ensemble of XGBoost models [32] that were trained on a dataset [33] based on the EXPObench library [4]. Here, a wind farm layout optimization problem using the FLORIS simulator...

    2020